Technical Report: Deprecated HuggingFace API Endpoint in Cloudflare AI Gateway

[jadmadi]

Example URL(s)

• https://developers.cloudflare.com/ai-gateway/usage/providers/huggingface/
• Error logs from Cloudflare AI Gateway dashboard showing: {"error": "https://api-inference.huggingface.co is no longer supported. Please use https://router.huggingface.co instead."}

Actual Behavior

The Cloudflare AI Gateway documentation for HuggingFace providers appears to be outdated. The documentation does not reflect that HuggingFace has deprecated api-inference.huggingface.co in favor of router.huggingface.co. As a result, users attempting to use HuggingFace models through the AI Gateway encounter errors indicating the old endpoint is no longer supported, even though the gateway URL format (https://gateway.ai.cloudflare.com/v1/{account_id}/{gateway_id}/huggingface/{model}) remains unchanged.

Expected Behavior

The documentation should be updated to:

1. Clearly state that the AI Gateway now routes to HuggingFace's updated router.huggingface.co endpoint internally.
2. Remove any references to the deprecated api-inference.huggingface.co endpoint.
3. Include a note about the transition and potential temporary errors during the migration period.
4. Provide updated examples or confirm that existing gateway URLs continue to work without user-side changes.

Additional information

[arunsathiya]

Ran into the same problem when trying to use HuggingFace. I connect to Z.ai directly, but would prefer to use via HuggingFace (which is in turn via Cloudflare AI Gateway)

[arunsathiya]

Quick update related to my last comment - For now, I decided to use Z.ai via AI Gateway as a custom provider, cutting off HuggingFace as a middleman. Ran into some hurdles along the way and shared some notes here:

• Custom provider documentation lacks clarity on endpoint path structure differences #27699

[jadmadi]

Custom Providers work flawlessly with HF, which makes it hard to understand why Cloudflare is taking weeks to fix a trivial issue with the default HF integration. Three weeks feels excessive.

At this point, I’m seriously considering abandoning default providers entirely and relying only on custom providers, simply to retain some control when things inevitably break.

[elithrar]

cc @palashgo

[github-actions]

Hey there, we've marked this pull request as stale because there's no recent activity on it. This label helps us identify long-standing issues in our repo (and not used to auto-close issues).